摘要
现有的绝大多数视觉传输系统致力于最小化接收端像素值的均方误差(MSE).但作为一种质量评价标准,很多情况下MSE 与人的视觉感官所得结果并不一致,而关于质量评价的研究表明图像梯度的结构可以更可靠地反映视觉信息.在此结论基础上本文介绍一种新的图像视频软传输方案,利用梯度数据承载视觉信息,以在无线视觉通讯中取得更好的视觉效果.此外本文为该方案的接收端设计了一种有效的图像重构技术,使之能够利用接收到的带噪梯度数据重构出高质量的图像.为此本文利用梯度统计特性进行自适应的分布建模,与全变分模型用一个固定的零均值拉普拉斯分布对梯度数据建模不同,本文利用图像块的非局部相似性从高度相关的图像内容中提取数据样本形成分布,每一梯度的期望、方差都自适应地进行估计.在此基础上本文提出非局部梯度稀疏正则化,根据逐像素估计的梯度统计特性对不同位置的梯度数据分别自适应地施加正则化约束.相应优化问题通过增广拉格朗日方法求解.实验结果表明本文所提技术比同类其他方法获得了更好的视觉效果.
Visual communication in wireless scenarios is quite challenging because the channel quality often varies unpredictably and dramatically with time and for different users. As a result, uncoded pseudo-analog transmission schemes like SoftCast have attracted a lot of research interests in the past few years due to the ability to handle channel variation effectively and achieve graceful quality transition in a very wide channel condition range. Such methods take advantage of linear transform, power allocation and dense modulation to attain efficient analog-like visual communication. A potential problem is that most of the existing visual communication systems aim to minimize mean square error ( MSE ) of pixels reconstructed at receivers. However, the quality metric MSE has long been criticized for not being consistent with perception of the human vision system. Grounded on recent image quality assessment studies showing that image gradients reflect visual information more reliably than pixel intensities, this paper presents a gradient based image/video SoftCast (G - Cast) scheme, which conveys visual information by transmitting the gradient data rather than DCT coefficients or pixel intensities, so as to pursue better perceptual quality for wireless visual communication. To be specific, G - Cast delivers gradient data that contains the majority of visual information of the transmitted image in the uncoded pseudo-analog way to cope with channel variation, and meanwhile sends out a very small portion of the low- frequency coefficients by a low bit rate with strong protection so as to provide global and regional luminance of the image. In practice, the transmitted gradient data are usually influenced by channel noise, which is commensurate with the channel signal-to-noise ratio ( CSNR ). To solve this problem, this paper develops an effective decoding method for the G - Cast receiver, which aims to reconstruct high quality image from the received noisy gradient data and a small number of low-frequency coefficients. For
作者
刘航帆
熊瑞勤
赵菁
李宏明
马思伟
高文
LIU Hang-Fan;XIONG Rui-Qin;ZHAO Jing;LI Hong-Ming;MA Si-Wei;GAO Wen(National Laboratory for Video Technology, Peking University, Beijing 100871;Center for Biomedical Image Computing and Analysis, University of Pennsylvania, Philadelphia, 19104 USA)
出处
《计算机学报》
EI
CSCD
北大核心
2019年第9期1905-1917,共13页
Chinese Journal of Computers
基金
国家基础研究计划（2015CB351800）
国家自然科学基金（61772041）
国家重点研发计划（2017YFB1002203）资助~~
关键词
软传输
视觉质量
图像梯度
自适应建模
稀疏
SoftCast
perceptual quality
image gradient
adaptive modeling
sparsity
